Mixture of Agents — Smarter Answers

Think of it like a panel of experts debating a tough question, then a lead editor writing the final report — that’s Mixture of Agents (MoA) in a nutshell.

---

What Is MoA?

MoA is a virtual model provider. Instead of one model, you get a package: several reference models plus one aggregator model.

When you pick an MoA preset, the aggregator handles your actual question and tool calls. But before it answers, the reference models run first and produce their own analyses. The aggregator then reads all of those and crafts the final response.

Simple version: multiple models brainstorm, one model summarizes.

---

When Should You Use MoA?

Use it when a task is hard and a single model might miss something. Good examples:

• Complex code migration plans
• Architecture design that needs multiple angles
• Contested technology choices

Important: MoA still runs inside Hermes’ normal agent loop — tool calls, multi-turn conversations, interruption recovery, and session history all work exactly as usual.

---

How to Choose an MoA Preset

MoA presets appear in every model selection screen, just like regular models.

CLI / terminal:

/model default --provider moa
/model review --provider moa

TUI: type /model, then pick a preset under the Mixture of Agents provider.

Desktop GUI: look for the MoA presets section in the model dropdown and click to switch.

---

The Quick Command: /moa

/moa is a one-shot shortcut — it applies the default MoA preset to only the current message, then automatically restores your original model.

/moa design and implement the migration plan for this flaky test cluster

Note: everything after /moa is treated as the prompt — no preset names are parsed. A bare /moa (no content) just shows usage info.

For a permanent switch, use the model selector (like /model), not /moa — so ordinary conversations never accidentally change your model.

---

What Happens Under the Hood?

Every time the main model is called, Hermes:

1. Finds your chosen preset;
2. Runs the reference models first — they only see the conversation’s text (no system prompts or tool call history), so it’s fast and cheap;
3. Attaches their outputs as private context for the aggregator;
4. Calls the aggregator with normal tool-calling ability;
5. The aggregator’s answer becomes the final answer;
6. If the aggregator wants to call a tool, Hermes executes it normally;
7. On the next turn, the whole process runs again based on the updated conversation.

---

Summary and Practical Tips

Core value: MoA uses multiple perspectives to improve answer quality while keeping Hermes’ full agent capabilities.

Practical advice:

• For everyday simple tasks, stick with a regular model — don’t waste MoA’s compute;
• When you hit a problem where “one model can’t figure it out,” try /moa for a quick taste;
• If a preset works well, switch to it permanently with /model so the whole session benefits.

Multiple models working together — that’s how AI stops going it alone. That’s the wisdom of MoA.